The monsoon is now being asked to give its own timetable
the monsoon is now being asked to give its own timetable: an independent assessment finds global ai weather models capture the south asian monsoon including the 2025 pause but still under predict the heaviest rain tails, a transfer learned cnn forecasts mumbai’s city level heavy rainfall days ahead, and a video style network trained on 85 years of reanalysis learns gridded monsoon rain from atmospheric fields.
the wet season now files its own arrival estimate.
Context
MAUSAM (arXiv, 2025) evaluates seven AI forecasting systems (FourCastNet, FourCastNet-SFNO, Pangu-Weather, GraphCast, Aurora, AIFS and GenCast) during the South Asian monsoon against weather stations, rain gauges and geostationary satellite imagery rather than reanalysis. It finds good skill across many variables but systematic errors, including underprediction of extreme precipitation tail events, and forecast errors 15 to 45% larger against observations than against reanalysis. AIFS is the most consistent of the seven.
The Mumbai paper (SSRN preprint, with a listing in Urban Climate, 2026) downscales GFS forecasts to 36 stations across Mumbai. Pre-training a CNN on the full rainfall record and fine-tuning on heavy rainfall events raised threat scores by 66.7 to 400% on lead day 1 and 82.35 to 377.8% on lead day 2 at the 95th percentile. On lead day 3 GFS fell to zero and the CNN with transfer learning reached 0.3 or more for most stations. The authors say the model detected all but one extreme event in the 2024 trial.
The gridded monsoon paper (arXiv, January 2026) treats pre-monsoon atmospheric and ocean fields from January to May as a video-like sequence and maps them to gridded rainfall for June to September with a CNN, trained on 85 years of ERA5 and IMD rainfall. The abstract gives no skill number.
The note says the assessment 'including the 2025 pause'. The MAUSAM abstract does not mention a pause, so that detail is unsupported here, not refuted. The Mumbai result is a statistical downscaling of a global model at 1 to 3 days, measured by threat scores at stations, and the percent gains start from a low base. A threat score of zero for GFS on day 3 makes the ratio large. The other forecast in the same area is a different model: an earlier arXiv paper predicts Mumbai rain only 6 and 12 hours ahead with a physics-informed ConvLSTM, so it is not the same work. The gridded monsoon model is a seasonal outlook, a different task from the daily forecasts of the other two.
Related work
- MAUSAM: An Observations-focused assessment of Global AI Weather Prediction Models During the South Asian Monsoon (arXiv) ↗Source for the seven model assessment against observations.
- Hyperlocal Extreme Rainfall Forecasts in Mumbai: Convolutional Neural Network-Transfer Learning-Based Downscaling Approach (SSRN) ↗Source for the threat score gains and the 2024 trial.
- Hyperlocal urban extreme rainfall forecast system: A convolutional neural network-transfer learning-based downscaling approach (Urban Climate, listing) ↗The journal listing of the same work, dated June 2026.
- A Spatio-Temporal Deep Learning Approach For High-Resolution Gridded Monsoon Prediction (arXiv) ↗Source for the 85 year, video-style gridded monsoon model.
- Enhanced Precision in Rainfall Forecasting for Mumbai (arXiv) ↗A different Mumbai paper: a physics-informed ConvLSTM for 6 and 12 hour forecasts.
Watch next
- Check what the MAUSAM authors say about the 2025 monsoon in the full text. For the Mumbai model, look for verification beyond the 2024 trial and against the operational forecast on false alarms, not only hits.
Sources
- MAUSAM: An Observations-focused assessment of Global AI Weather Prediction Models During the South Asian Monsoon (arXiv)arxiv.org
- Hyperlocal Extreme Rainfall Forecasts in Mumbai: Convolutional Neural Network-Transfer Learning-Based Downscaling Approach (SSRN)papers.ssrn.com
- A Spatio-Temporal Deep Learning Approach For High-Resolution Gridded Monsoon Prediction (arXiv)arxiv.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 4 October 2026 at 02:16 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →